At Apple, we believe that protecting user privacy is a fundamental human right. To achieve this, we’ve developed new techniques that enable us to understand usage trends and aggregated insights without revealing individual behavior or unique content. These advancements will be used across various Apple Intelligence features.
One area where these new techniques are applied is in Genmoji, our emoji-based input feature. By using differential privacy, we can identify popular prompts and patterns without linking them to individual users. This ensures that only commonly used prompts reach Apple’s servers, while user-specific information remains private.
For longer text inputs, like email messages or summaries, synthetic data generation is used to improve model performance. Synthetic data mimics real-user interactions but doesn’t contain any actual content. It’s created using large language models and iteratively refined through curation steps.
These new techniques offer a balance between improving AI features and protecting user privacy. As we continue to advance the state of machine learning and AI, our commitment to user privacy remains unwavering.
Source: https://machinelearning.apple.com/research/differential-privacy-aggregate-trends